Homogenization of daily temperature data of the five principal stations in the Netherlands - version 2.0

This report presents version 2.0 of the homogenization of records of daily minimum temperature TN, daily maximum temperature TX and daily mean temperature TG at the five principal stations in the Netherlands. These historical records – starting at the beginning of the 20th century – have been influenced over time by station relocations and changes in measurement equipment. For four of the stations – Den Helder/De Kooy, Groningen/Eelde, Maastricht/Beek, and Vlissingen/Souburg (the H4 stations)—parallel (simultaneous) measurements at the old and the new locations were made over periods of 4–10 years to support accurate adjustment. At De Bilt, only the effect of a screen change in 1950 was measured in parallel; the impact of the near-simultaneous  relocation was not directly recorded.

Our approach to homogenization rests on the following premises: (a) homogenization is applied cautiously and only when supported by evidence, focusing on adjusting for known and substantial changes in instrumentation or location. (b) gradual environmental changes, like urbanization, are not adjusted for due to their complexity. (c) homogenized data are versioned and stored alongside raw data for transparency and future updates. (d) robust methods are preferred, and automatic breakpoint detection is used only for quality control – not for undocumented changes – to avoid introducing new errors. We further note that daily maximum temperatures may be affected by turbulent eddies during warm sunny days. Daily minimum temperatures may be extremely sensitive to small breezes during stable nights. Even the existence of high-quality parallel measurements cannot fully account for these effects.

This updated version of the homogenization was developed for three main reasons:
1. Recent research (de Valk and Brandsma, 2023) has shown that for the H4 stations where parallel measurements are available, daily temperature adjustments to
account for instrument relocation can be improved by including other weather variables – like wind, humidity and cloudiness – that influence temperature differences between sites.
2. For De Bilt, where parallel measurements cannot be used for homogenization, a more robust adjustment can be obtained by a careful choice of the reference data to be used for this purpose (Dijkstra et al., 2022); in particular, the use of homogenized data instead of raw measurements and careful selection of station(s) and calibration time-intervals.

3. There is growing demand for making both the original and adjusted datasets more transparently available, allowing users to assess how homogenization affects their analyses.
For the H4 stations, two methods are compared: an updated version of the original Quantile Delta Mapping (QDM) technique and a newer method using Generalized Additive Models (GAMs). The latter enables the use of measurements of additional weather variables to improve the accuracy of the temperature estimates for individual days. For De Bilt, where no suitable parallel dataset exists, an improved QDM method is chosen, because we cannot show that GAMs can offer benefits. The QDM method uses longer calibration periods than in version 1.0 (15 years before and after breakpoints) and the average of the homogenized time series of two inland stations (Eelde and Maastricht) as reference. This leads to more precise and robust adjustments for the change in screen combined with the relocation in 1950-1951.

Key findings and conclusions:
1. The homogenization helps to align trends across stations.
2. For the H4 stations De Kooy, Eelde, Vlissingen, and Beek/Maastricht Airport, the updated version leads to more realistic day-to-day adjustments (reductions in mean square error between 6% and 76%). However, the long-term trends differ little from those of version 1.0. This suggests that the earlier method was already fairly reliable for these stations.
3. For De Bilt, the adjusted version 2.0 method improves precision and robustness when compared to the previous version, with the caveat that validation is limited by the absence of suitable parallel measurements.

4. For De Bilt, differences in long-term trends between the versions and between version 2.0 and the unadjusted data are substantial only for climate indices indicating heat, such as the number of heatwaves. The number of heatwaves in 1901-1950 is now estimated to be 14, twice as many as in the previous version. However, even for these indices, the long-term trends from both versions cannot be conclusively distinguished, because of their high uncertainty due to natural year-to-year fluctuations. For other climate indices, the differences are minor.
5. Homogenization can introduce errors if the models used are not calibrated with sufficient precision, for example when calibrated on a dataset that is too small. We find that the version 2.0 homogenization has been calibrated with sufficient precision; the calibration has little impact on the overall precision of long-term climate trends, which is determined mainly by the natural year-to-year variability.

Checking for inhomogeneity of measurement records and – where needed – homogenization helps to ensure the temporal and spatial consistency of the national climate
records, which is essential for the analysis of the climate of the Netherlands and its change over time.